Stale Worktree Cleanup
SkillDatabases & dataLets your agent find and delete old git worktrees and drop orphaned Postgres databases.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Stale Worktree Cleanup skill
About this capability
Use when cleaning up old git worktrees, removing worktrees whose branches have merged or been abandoned, or dropping orphaned compdev_* Postgres databases. Triggers on "clean up worktrees", "delete stale worktrees", "worktrees piling up", "orphaned databases", "remove unused worktree".
What this skill tells your AI
The instructions your AI receives, as published by trycompai/comp in .agents/skills/stale-worktree-cleanup/SKILL.md and read by ahel’s review.
Overview
This repo's .githooks/post-checkout creates a per-worktree Postgres database (compdev_<slug>) on every git worktree add. Git has no pre-worktree-remove hook, so dead databases pile up over time. This skill defines a safe, reversible process to reap both the worktree directories and their dangling databases together.
Core principle: never delete work the user might still want. Classify first, ask second, remove last.
When to Use
- User says "clean up worktrees", "delete stale worktrees", "remove old worktrees", "worktrees piling up"
- User mentions orphaned
compdev_*DBs or running out of disk space from deadnode_modules - Starting a new feature and noticing many old worktree dirs
When NOT to Use
- User wants to remove ONE specific worktree (just run
git worktree remove <path>+ drop its DB directly — don't load the whole process) - User is actively working in a worktree (never touch active work)
Process
Step 1 — Inventory
Run these four commands (parallel-safe) and capture the output:
git worktree list --porcelain
gh pr list --author @me --state all --limit 50 --json headRefName,state,url,number
git branch --no-color | cat
Then query the DB for compdev_* databases. The management URL is the main worktree's DATABASE_URL with the database path swapped to postgres:
bun -e '
import { Client } from "pg";
const raw = require("fs").readFileSync("packages/db/.env", "utf8");
const url = raw.match(/^DATABASE_URL=(.*)$/m)[1].replace(/^["\x27]|["\x27]$/g, "");
const mgmt = url.replace(/\/[^/?]+(\?|$)/, "/postgres$1");
const c = new Client({ connectionString: mgmt });
await c.connect();
const r = await c.query("SELECT datname FROM pg_database WHERE datname LIKE \x27compdev\\\\_%\x27 ORDER BY 1");
for (const row of r.rows) console.log(row.datname);
await c.end();
'
Step 2 — Classify each worktree
For each worktree (skip the main one):
| Signal | Classification |
|---|---|
| Branch merged to main AND clean working tree AND no unpushed commits | safe |
PR is CLOSED (not merged) | needs-confirm |
| Uncommitted changes OR unpushed commits | needs-confirm |
| No matching PR, no merge, has local commits | keep (user may still be working on it) |
| Is the main worktree | skip |
Gather per worktree:
cd <path> && git status --porcelain | wc -l— uncommitted changes countcd <path> && git log @{upstream}..HEAD --oneline 2>/dev/null | wc -l— unpushed commits (0 if no upstream)- Branch → PR lookup from step 1
Step 3 — Present to user
Show a table and explicit recommendations. Example:
Path Branch PR state Changes Recommendation
.worktrees/sale-45-… mariano/sale-45-… MERGED 0 / 0 ✅ safe to remove
.worktrees/old-experiment mariano/scratch — 3 / 0 ⚠ uncommitted — confirm first
.worktrees/worktree-env-auto-link mariano/worktree-env-auto-link OPEN 0 / 0 ⏳ keep (PR open)
Orphan databases (no worktree dir):
compdev_abandoned_feature
compdev_old_migration_test
Then ask: "Remove the items marked ✅? Confirm by listing anything you want me to also nuke from the ⚠ / ⏳ set."
Step 4 — Remove confirmed items
For each worktree the user approved:
# 1. Remove the worktree dir (use --force only if user confirmed dirty-removal)
git worktree remove "<path>" # clean case
git worktree remove --force "<path>" # only after explicit user OK
# 2. Derive the slug and drop the database
slug=$(basename "<path>" | tr '[:upper:]' '[:lower:]' | tr '-' '_' | tr -cd 'a-z0-9_')
bun -e '
import { Client } from "pg";
const raw = require("fs").readFileSync("packages/db/.env", "utf8");
const url = raw.match(/^DATABASE_URL=(.*)$/m)[1].replace(/^["\x27]|["\x27]$/g, "");
const mgmt = url.replace(/\/[^/?]+(\?|$)/, "/postgres$1");
const c = new Client({ connectionString: mgmt });
await c.connect();
await c.query(`DROP DATABASE IF EXISTS "compdev_'"$slug"'"`);
await c.end();
console.log("dropped compdev_'"$slug"'");
'
For orphan databases (no matching worktree dir), just drop them — no worktree to remove.
Do NOT delete the local branch unless the user explicitly asked. Branches can be recreated from origin cheaply; worktrees cannot.
Step 5 — Verify and report
git worktree list
# then re-run the compdev_* query from step 1
Report back:
- Worktrees removed (paths)
- Databases dropped (names)
- Anything skipped and why
- What's left (still-active worktrees)
Safety Rules
- Never remove the main worktree. Its path is always the first line of
git worktree list --porcelain. - Never
--forcewithout confirmation. Dirty worktrees can contain un-stashed work. - Never
DROP DATABASEon anything not matching^compdev_[a-z0-9_]+$. Never dropcomp,postgres, or any production-looking name. - Never delete a local branch as part of cleanup unless the user explicitly asks. Orphaned branches are cheap; lost work isn't.
- Never run this inside a worktree you're about to delete.
cdto the main worktree first.
Common Mistakes
| Mistake | Fix |
|---|---|
| Dropping the DB but leaving the worktree dir | Run git worktree prune then git worktree remove |
| Removing the worktree but leaving the DB (accumulates orphans) | Always do both in the same pass |
| Using hyphens in the DB name | The hook slug rule is tr '-' '_' — always underscores |
| Running from inside a doomed worktree | cd to the main worktree before starting the process |
Using gh pr list on a branch with no PR | Missing data is not "abandoned" — needs needs-confirm classification |
Red Flags
If any of these show up mid-process, stop and ask the user:
- A worktree has unpushed commits AND no PR → might be unreleased work
- The classification returned >10 "safe to remove" items → unusual volume, double-check
git worktree removeerrors with "working tree is not clean" → never retry with--forcewithout explicit consent- A
compdev_*name has weird characters or unexpected format → don't drop
Quick Reference
# List everything
git worktree list --porcelain
gh pr list --author @me --state all --limit 50 --json headRefName,state
# Per-worktree inspection
git -C <path> status --porcelain
git -C <path> log @{upstream}..HEAD --oneline 2>/dev/null
# Clean removal
git worktree remove <path>
# Dirty removal (requires user confirmation first)
git worktree remove --force <path>
# Database drop (run from main worktree)
# See the bun -e snippets above for the exact invocation
Signals
- GitHub stars
- 2k
- Forks
- 414
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
stale-worktree-cleanup- Source
- github.com/trycompai/comp